AWS Summit Series 2019 – Santa Clara: Keynote featuring Werner Vogels


Please welcome vice president and chief
technology officer amazon.com Dr. Verner Vogels good morning Santa Clara we did come
shake to the left to the right you know meet people next to you no I won’t do
that to you good morning as always we’re very very
proud that so many of you are willing to come out here today to listen to where
we are with the Amazon Web Services yes pioneers in the whole cloud world
we’ve developed so many services over time where each of you sort of should
have the right tools to build the applications that you want to build
something that we’ve always said this that we couldn’t have come here without
you meaning that your feedback to us and
working very closely with you our customers to actually build our services
and to develop a world map is unique about 95% of the features and services
that we’ve delivered until now have come in direct feedback from you yeah so and
it’s most important because I think and if we would have been building the tools
for development as old companies who have done that
we would have built the tools that you were using five years ago and there’s
not we want to build together if you the tools that you need to be ready for 2025
yeah so that niggle in lockstep because the development is changing is rather
than changing both the ways in which we architect the operational models the
security postures the way we use data all these things are changing radically
and as such we really rely on you and working closely together with you to
actually build a web map for that say due to modern application development so
after spending a little bit of time on the business update I’ll go deeper into
so the patterns that we’ve seen arriving with most of our customers have other
sort of modern development and what are the kind of things that we’ve built to
help you build really two modern applications so yeah six and half
thousand of you are out here thank you very much I think you know you all have
busy day and a switch it’s a it’s pretty humbling
that so many of you come out here to listen to us as always I consider these
events to be educational events and not sales events that’s not why we’re here
those 50 well over 50 deep dive technical sessions is really the meat of
this gathering where you can really get the glow down we need all the details on
whatever pieces of AWS you’re most interested in of course you know it’s a
part and network is continuing to to grow and if you go out into the expo
hall you’ll find many of our partners having their stands there go talk to
them go hang out with them go listen what kind of pieces they’ve developed
for AWS I’ve always said that AWS is so much more than just AWS you know without
all our partners that are building whether it’s operational tools or
whether it’s ice fees or software all of them extend AWS in a manner that makes
it extremely rich yeah whether you’re integrating Twilio into your application
or stripe or any of the other parties that we have you know that really made
State we as platforms so much more richer then all the services that we’ve
built ourselves already so let’s take a quick look at where we are with the
business yeah at this moment based on the core fourth quarter results from
from last year we went we are at a close to 30 billion dollar Redmond rate and
you know on that base a forty five percent growth year over year is pretty
astonishing and if I believe if I look at some of the past 13 years of AWS I
think it is the speed at which we’ve grown as has
really been the biggest challenging thing you know we’re continuously
changing IT landscape and working together with you to really grow really
fast built very vivid the tools that you need I think is 45 percent growth year
of here also shows that we are doing the right thing by working closely together
with you building a set of tools that you really can use
to build your next generation applications it has we’re very fortunate
that this has resulted in literally millions of businesses running on AWS
and an active customer is something we be considered to be and non Amazon
entity that is that has been active in the past 30 days literally millions
millions of businesses winning on AWS and maybe that is startups and actually
a fine startups to be a bit of a misnomer these days because many of
these names on this slide here are household names Emilia with us lifts or
whether it’s uber or whether it is Dropbox over that’s air being be a
little bit slack or Pinterest maybe these companies were at one moment
a start-up within my eyes what I would rather cover this internet skill
companies that I really focused on being internet first and I rather lifting all
email AWS robin hood from here from the neighborhood actually is building his
mobile application for for trading five markets around here
you know they build a highly scalable analytics platform on AWS that allows
them to go allowed them to go from zero to a hundred million revenue in just
over 14 months and of course it’s not just not just the startups maybe in the
earlier days this was sort of the idea behind the AWS to really serve these
companies that wanted to reach internet scale I think enterprises have figured
out that this is way too good for them as well and very closely working with
them building a whole new set of capabilities yeah
Expedia is moving all in into AWS capital wang built a digital leading
banking platform on AWS and today was a great announcement
standard bank one of the largest lenders in in South in in Africa it’s moving all
of the infrastructure over to AWS and another announcement today will set for
box wagon is collaborating with AWS to build an industrial platform for
managing the efficiency in all of their plants and as you can see you know it’s
a wide variety of enterprises making use of a Tobias there’s almost not
ethical where there’s not companies that have decided to go all in on to AWS and
maybe that this enterprises but of course also in the public sector
thousands of agencies around the world making use of AWS because most of the
government’s you know every dollar or every euro he can save actually is is
money that you can put towards programs that really matter for your citizens and
so whether that’s for example the UK Ministry of Justice who have built a
whole pipeline of services that help law enforcement and prisons and and and all
sorts of other activities around there that have very high sensitivity in terms
of privacy and security city of Los Angeles build a whole security system in
and around all of their of the city’s departments in gathering their data and
analyzing for security risks and so across the board long the nonprofit
organizations that salsa government agency making use of AWS we do not gonna
do that with other partners especially I think in the in the enterprise world
many of these organizations already existed for a long time is global system
integrators like a center in Capgemini and others but also new born in the
cloud system integrates like like second watch and so many of them are the ones
are likely really helping our customers move on to AWS especially those that
have for example challenging environments for example based on CSAP
and things like that and most of these partners have great competencies in
actually helping you get there if you’re a nice fee or software as a service
vendor you own a WS why because your customers are there your customers will
demand that you there most of these fees they’ll be moving to a to a
software-as-a-service model in terms of delivery and whether it’s it’s a Keanu
with Adobe which was info informatica or Salesforce or workday Splunk
all of them have moved to software as a service model to deliver their
functionality on top of AWS now I’m always fortunate in these events
to have great guest speakers and so a first speaker was actually transforming
his company into a cloud first business so f5 which is a provider in application
security services became working with AWS about six years ago and using the
marketplace and becoming a networking security APM partner come competency and
so since joining f5 as president and CEO in 2017 concerns warlock do Co you know
has focused on accelerating this effort from moving from the traditional company
that that they were into a solutions and services company delivered food cloud
now to be more about this I would like to welcome from swatter States well I
would like to begin my story today by what most will assume is the end success
I believe that success can kill a company the classic definition of
success assumes that we have reached the highest of heights and that therefore
what we have accomplished must be protected that is a mindset that leads
us to a very dangerous place that I call the status quo the status quo is in fact
the biggest threat to our companies to our cultures to our personal growth
because it is a familiar friend who is hard to resist
and even harder to say goodbye to so I want to share with you some of the
symptoms that you see in an organization that is under the spell of the status
quo comfortable bureaucracy settles in you accept that it takes a long time to
do anything your customers warn you that things are changing but they continue to
buy for the time being in the form of institutional arrogance sets in where
how we do things takes over becomes more important than
curiosity and invention and while these are signs of the status quo and
organization I believe that there is one fundamental quality that holds any of us
apart from succumbing to the status quo it’s a
quality that no money can buy you either have it or you don’t
and that quality is the drive to reimagine it’s the belief that your
first success can not be your last it’s the courage to challenge our own formula
for success because we know it can be done better and so I am here today to
tell you the story of f5s rejection of the status quo a feat we have undertaken
more than once in our history and what it took for us to reimagine again now
for f5 over the last 10 years the status quo looked like this it was a company
famous for its load balancers but also obsessed with a hardware business model
it was all suffering application security and delivery services to the
top mission critical workloads in a day Center but leaving tens of millions of
other workloads unattended it was a loyal base of net ops users
inside 25,000 enterprise customers but very little to offer to the growing
DevOps communities inside the same organizations now as they say what got
us here won’t get us there out of five we have a mission we want to provide
enterprise grade application services for every app anywhere and the only way
to get there is through the cloud now as it does for you the cloud requires a
continuous transformation of our business for f5 that meant significant
important but painful decisions we had to completely redefine our customer
personas who we aim to serve and how we had to make significant shift in where
and how we invest our resources and relook at the behaviors we promote in
our own organization and we also had to create startups carve out startups from
within f5 with a very clear new charter to
disrupt the status quo the result of this is an f5 that is now offering easy
to consume friction-free application services consistent
application security for every workload across every environment and a company
that is finally bridging the divide between net ops and DevOps by joining
forces with nginx the leading apple open source application delivery platform we
can now offer enough effective controls to satisfy the CIO but also enough
freedom to for application developers and it should
be no surprise that for a company committed to disrupting its own status
quo f5 chose AWS we built our cloud services platform leveraging the breadth
and depth of AWS infrastructure services we do storage and compute of course but
also caching an identity database and even server lists we worked with the AWS
SAS factory team to transform our own development process and build and
deliver new services 50% faster we also leverage the AWS marketplace it allowed
us to build digital procurement on a global basis to companies ranging from
startups to Fortune 500 to our own channel partners 12-month faster I’ll
say that again 12-month faster than we would have on our own and leveraging
that the built-in metering features and digital commerce enabled by the AWS
marketplace millions of f5 customers can now try and subscribe to our five
services in minutes the result of this work is the next critical step in our
reinvention f5 cloud services I am pleased to announce that we are
launching today on AWS f5 cloud services f5 cloud services is a family of cloud
native solutions designed for enhanced application delivery security and
insight and it’s immediately available for our customers and through our
channel partners on the AWS marketplace it starts with our DNS cloud service and
a preview of our global server load balancing service available for use in
AWS or in hybrid cloud environments and later the spring we will be
delivering even more at five enterprise-grade sass capabilities
including security services designed to protect applications from both existing
threats and emerging threats thank you thank you
there must be somebody from f5 in the room here the best part about all this
though is we are just getting started I don’t believe we find an endpoint in
success the spirit of the summit each of us is asked to consider what can we do
differently and I know this inspection can be painful but I can also share with
you what it feels like when you have rejected the status quo you are restless
for more the risk-taking feels less risky and new ideas are courageously
surfaced every day I know that is what it feels like an f5 now and the
opportunity for all of us here today to invent to grow and to break away from
the status quo thank you Thank You Francois yeah that’s why we
built AWS to help everybody break the status quo yeah because in the status
quo your vendors were in charge not you and one of the biggest things that we
tried to do we built AWS is to take all not pieces of the motto of Amazon the
retailer to be the Earth’s most customer centric company so how do you do that as
an IT provider how do you become the world’s best customer centric IT
provider is by putting your customers in charge you’re in charge of our world map
but also the economic models that we put in place we’re really there to put you
in control instead of earth as a provider and so I
go you’ll continue to be the Earth’s most customer centric IT provider to do
that we need to move away from the Mobile’s that we had in the past where
we as a technology provider would give you everything in the kitchen sink and
tell you and this is how you shall use it now think of the new world everybody
knows that you need different tools for the different jobs and their search
we’ve been really focusing on making sure that you have choice and building
the broadest and deepest platform for you as builders today so you can pick
exactly those tools you want to build well maybe maybe in the past we you were
building a house it was sort of a prefab thing and it was sit there it couldn’t
do anything about it maybe there’s two or three of these
houses that you could choose from but if you really want to build unique houses
you need to have unique tools to really build exactly the house that you want to
have and I think that’s really where we are today so well over I think a
homeless 65 different services in AWS right now and that’s continue growing
and whether that is in analytics or IOT or machine learning or mobile services
or the blockchain technology or DevOps the days from where
AWS was just infrastructure-as-a-service yeah
compute storage databases security those days are long gone
mostly because you’ve been asking us once we solve that say the heavy lifting
in the infrastructure to start solving the heavy lifting the other pieces of
heavy lifting that you still have and that’s why we continue to roll out these
new services based on your feedback what you need and let’s pick on a few of
these if you look at databases now we have 14 different database services and
so of course relational still plays a quite important role because many of you
actually have real need for relational databases but sometimes also you using
standard off-the-shelf packages that only run if you have video relational
database backing it up and so but what we see more and more and especially if I
move to micro services where the components become much smaller and they
really make use of purpose-built databases to meet exactly the needs of
their application rather than is key-value already there’s graphs or
whether it’s a ledger you know all of them have unique capabilities that we’re
using today instead of using the relational database as a hammer that you
can use for everything we are moving to really specific high performance highly
reliable managed purpose-built databases the same goes for security security a
hundred and sixteen of our services have encryption enabled in them 52 of him you
can bring your own keys and I’ll be talking more about it later
with respecting in terms of security but encryption is becoming the most
important tool you have to make sure that you’re the only one who has access
to your data and nobody else and whether there’s also in combination with all the
the compliances and all the certifications that we’ve achieved or
all the innovation that we’re doing under the covers to build new automation
tools for you where you can actually protect yourself and I think automation
in security plays are very important well and we’ll get back get more into
that in a bit the same goes to storage it’s not enough
to just have a volume service yeah a lot for block sales you need to have
different variations in there because all of you have different types of
workloads you know and so you can where you can actually really tweak your
volumes exactly to meet the requirements that you have and then singles for the
different types of oil object storage as well import it in all of this is that
you can pick exactly that tool that you need the same goes for for example
instance types in in the past maybe he was been you would be stuck with this
particular type of server and you had to develop your software for it these days
however you develop your software and then you go look for what is the best
instance type that actually matches what I need to do in my applications and
there are also going to storage of course and I think still sort of the
ninth world wonder in terms of digital sense is Amazon s3 now it’s it’s the
first service that we launched in AWS is now 18 years old and customers are
routinely processing exabytes of data when it comes to sv and so whether it’s
all the mechanisms that we developed the way you can automatically move between
different storage classes or the security capabilities that we put in
there all the object level controls that you have it’s been an amazing feat to
see how IC has evolved over time and even to the point that you can use ratio
spectrum to your basically your data warehouse to point to your exabyte data
set that lives in s3 and just to understand the data warehouse queries
over it it’s pretty spectacular now one of the stores classes of course in a3 is
we Nestle pleasure which is there to do long-term archiving and we we announced
Glacier deep archive at reinvent and I’m very happy to announce that that one is
now general available for everyone to use reported pile of glacier deep archives
is that it comes at a cost point of not even a tenth of a cent per gigabyte per
month yeah no more no more glacier no more tapes and it gets the same level of
durability that you see in s3 and so this has a whole set of use cases that
is pretty spectacular one of them that I like because of my own history in in
health care is the fact that in the past we had MRI and cat scan images they
would be archived of film where the hospitals have a requirement that these
data should be kept around for at least 30 years so instead of keeping the
digital formats they basically printed it on film and then had radiologists
later compare film instead of the digital images mostly because I couldn’t
afford it and there were no storage systems available that would actually
sort of have this level of durability over such a long period of time and so
at these particular cost points yeah hospitals and others can now start
storing huge amounts of data at a very low cost but be guarantee that the
durability will be there for for decades to come
so if all of these capabilities you know we definitely see a move to newer types
of applications so if I look at what most of our customers are doing it’s
sort of interesting to see that amazon.com the retailer actually went
through this phase five to ten years ahead of that we needed to reach scale
and reliability performance and security that most of our customers now I started
to get confronted with so if you look at the how development transformed at
amazon.com it really went from what we know is sort of a monolithic application
over to what’s now a hole deep microservices and environment and so you
can talk a bit more about micro services in in a minute but it was in a digital
business like like Amazon experimentation and fast experimentation
is crucial continuously experimenting and mama lives are not
good for that they have a very hard time to have this wrong big piece of software
where many different teams have to work on together really really fast moving
innovator and an experimenter doesn’t really work and so next to scaling
issues and all the technology issues that we had been running a monolith we
really wanted to break up all of that into a manner such that we could move
faster and in the joking term that we use was that of two pizza teams so teams
that a service has a team associated with it
that is really responsible for that service and completely it’s full and
total ownership also over there web map now so these teams live by what I used
to call you build it you run it and that sort of the first days of what we now
know as DevOps so it was important for us because now that we’ve built this
sort of decentralized environment not only from an architectural point of view
but also from an organizational point of view we could move really fast we could
start making new versions of some of these micro services and experiment with
them in a much easier way that we ever could have done with a monolith and a
good example there is actually coming back to s3 when we launched a suite 13
years ago we had eight separate micro services that made up se some that they
put in order to get and some of them did the scanning storage and maintaining the
index so only eight of those but what we knew on day one when we were building a
suite is that that would not be the architecture that we would be willing to
see four or five years later most of if every order of magnitude growth you have
to sort of revisit your architecture now if this would have been a massive moment
if this would have been a nightmare then at one moment you would get an email
from Amazon saying like oh we’re taking SC of wine on Friday night from 10:00 to
midnight to build a new version that would not be a good plan with it and so
as such you need to be able to evolve your software while your customers are
still running so now se is well over 250 five different
distributed micro-services if also to new capabilities that we’ve been
building in overtime and also lessons that we learned yeah we we learned that
hardware no matter how high end it is trails at times and do really weird
things like incorrect a bit flips in Rama that suddenly happen and even
though you know we so you have you built a micro services that take that one
particular job really well and the thing with micro services is that many of
these decomposed building blocks that you have out of human olive actually
some of them many of them have very different scaling and reliability
requirements and we’ll get back to that so if I look at sort of when we when
Amazon as well as our customers go through his move from from monolith to
micro services what’s the impact on the way that we develop our software and
that we operate it so let me go through some of these different phases in there
so first of all we look at this these architectural patterns that the move
from monolith to micro services is it’s probably one of the biggest
architectural changes that I’m seeing in the past years that the most important
thing and it so let me tell his story there about actually about bad amazon
the retailer so when we broke up the first monolith we had we had three very
large data sets so customers items as a catalog and orders and basically what we
taken sort of business for Jake moved it away put it next to the databases and we
had these three very large services left and one of them was the customer master
service basically all code that operated on the customer master database now
we’ve learned over time pretty quickly that that was a mistake
we done a data driven decomposition of our system and we should have done a
functional decomposition because in that customer master service you would have
one component that would basically be the recognized customer service a login
service let’s call it actor and in that same piece of software would also see
the address book service that was only needed when you do a check out yet let’s
login it’s almost hit on every page so now the
whole component needs to scale at the scale of that smallest component that
sits there press that is whole software component is whole blob has access to
both the credential store as well as the others bookstore which is almost a
violation of security properties and so really being able to decompose into the
smallest building blocks that you can imagine and then have each of those
scale along the dimensions that they need to scale and so the login service
just by itself can scale to mend asleep without impacting the others book sales
now ten years ago or it was at 20 years ago when we started going through this
process and learning about it I wish we had containers we didn’t yeah
and so if you look at sort of the different types of compute available for
you to support all of this instances VMs virtual machines containers and lambda
all play an important role and there’s a clear shift happening over time away
from instances into sort of more service development but instances will be around
for a very long time yeah so at this moment we have a hundred and eighty
instance types for you and very that is sort of well as burst capability or
general purpose or memory intensive or disk intensive or you to memory blobs if
you want to when your SP Hana systems all of those all of those capabilities
need to be available because of so many different workloads that are available
and so where you can pick among load whatever instance type you really need
to support the application that you’re running now which if a hostess for is
whether we can also reduce cost further and so happy to announce today that the
AMD AMD based instances are generally available both in terms of both in the m and the our category so
that is just general purpose as well as memory tens of workloads and so they’re
all based on the aimed I am the epic 7000 processor and basically they have
exactly the same numbering the same family evolution as that the
general-purpose Amazon are I have the Intel based once so it can immediately
start switching between ball in the other the advantage of the AMD ones is
that they’re about 10 percent lower cost than the previous instances we had if we
look at containers that clearly is the point where I see most of our customers
that are moving to a micro services or vomit are actually making use of at this
moment yeah and there’s a real rapid evolution in an amount of containers
happening because it makes it so easy to actually build this micro services
environment on top of this and we have really lots of customers that are
actually really experimenting or building real production systems using
containers so think about McDonald’s yeah it’s by any chance the world’s
largest restaurant chain with 37,000 locations around the world serving 64
million people a day so they built a home delivery system they did this in
four months using ECS the Amazon is yesterday elastic container service and
they serve 20,000 orders a second out of that micro services environment using
containers typically latencies are 100 milliseconds so this is an amazingly
scalable environment that really has all the components all the way built other
components to reach this massive scale also for the whole API system around
them such that they can actually integrate with partners such as ubereats
and then almost a pretty impressive development of of back-end services now
if you look at the capabilities available on on AWS there’s different
places where you need to make decisions you
so are you going to use the elastic container service or are you going to
use the lesser container service for coup Bonita so ECS or yukia’s those are
the two choices you have at the orchestration level at the compute level
underneath there underneath the containers you have a choice whether you
want to manage those clusters yourself or when you want to make use of Fargate
which turns your container service into a surplus container service where you
only have to worry about sort of building the software that has to win
your containers and not worry about the infrastructure anymore and of course
with all of that you need to have a container registry service that needs to
be highly scalable we have some customers that pull the same image 4,000
times into different tasks and so security and scale and reliability of
the container service is crucial in all of this the harder to make choices
between the different container services is is more or less you know it’s a
choice you make often between highly opinionated systems or ones that
actually give you way more flexibility so whether it is when you value
simplicity over flexibility and so if you look at the EGS that’s clearly where
I think simplicity rules it’s a highly opinionated service about how to build
container based applications it’s very deep integration with each and every one
of the other AWS services and whether that’s a OB or cloud watch or guard duty
all the integrations there are are crucial and so especially when it comes
to out of scaling and scaling over multiple availability zones this is
crucial and so many of our customers are making
use of ECS because of this deep integration into AWS and most of these
are customers that really start building their first container systems on AWS
itself ets however is much more flexibility oriented although we’re
deeply engaged with the if the open source community
I’m CUBAN eaters and we are hold a present upstream that means that we push
all of our changes or all of the integrations that we do in AWS into the
general repositories first and get them accepted there before we start launching
this in eks itself again we’re working on getting deeper integration into the
LMS platform but it also allows you often to already start developing this
on your laptop or maybe on-premise and then start moving your container service
over to AWS and I see ku Benitez mostly happening but many of our customers who
are looking to migrate into the cloud where they really start building things
in their own environment with the idea that they will be able to move this over
to the cloud whenever they get to that particular point in all of that I’ve
always been definitely in the early days of container based systems I was kind of
surprised there about the willingness of everyone to manage again resources at
the lowest level because if you think about containers really think about the
applications you want to build you don’t want to manage servers or instances
clusters underneath there so we built a to be as far gate to take away all of
the heavy lifting that comes with winning container systems because really
you don’t really need to manage those clusters there’s no value in that if you
really want to finish focus just purely on building business logic and so
whether the business logic runs in a container or whether you actually make
use of truly service environment like lunda that’s available for each and
every one of you so if you think about sort of the continuing from from
instances to containers and to lambda it is clear that there is a massive drive
happening there and many of our customers especially those at a cloud
first and thinking about building new applications all start off with service
today and why because the productivity is much higher and you don’t have to
think all these other pieces that you have to
do around sort of provisioning infrastructure winning things that were
multiple disease managing your security posture things like that many of these
are all taken care of by lambda itself and of course we continue to innovate
looking at how you are building these service applications and it’s very
interesting because this is a continuing and we continue to work with you because
this is such a new world surface that we need to make sure we building the right
tools for you and so layers has become one of the important components on one
hand to make sure that you don’t have to upload redundant pieces of code you can
share this piece of code between different applications or version them
and things like that but also it’s an ability for you to actually really get
one of your own application runtime so your language runtimes and it’s an
integrating there as in a lambda layer so you can learn any programming
language that you want to read now every C custom is building pretty extensive
things if you look at this this is home away it’s a company by Expedia and so
what they have tell you about six million images are being uploaded there
this is say for clarification home brokerage service people upload about
six million images each month with all these images need to be transformed into
what into standard sized images and you know firm nails and also tended need to
be pushed from machine learning to see whether these images are appropriate and
all these kind of things and as you can see in the whole architecture no servers
yeah everything is a combination between lambda and other service components like
dynamodb an SV and Kinesis it’s a pretty this is a pretty common architecture
today where there are no servers in this picture anymore you literally have
hundreds of thousands of our customers that are all using lumber and the most
amazing thing happens in all of this we think about solid building new
technologies it’s often the young tech not a technology startups that are sort
of adopting technology first but what we see with
with Linda is that actually enterprises jumping on board immediately and why
because it is makes it so easy to only have to pay for those resources that
you’ve really used very effective management mechanism but also in create
so much greater productivity with your developers that that’s really something
that as an enterprise if you want to move fast are really concerned about so
whether that is for example a company like Capital One they migrated billions
of mainframe transactions into a system with DynamoDB lambda and other AWS
services basically completely eliminating the mainframe instead going
over to another container or another image based system now they moved over
completely did jumped all of those steps and actually started using a DBS lambda
to replace their mainframe with well if all of these different components coming
together are you might have different languages might have different types of
applications and all of them are sort of running in this distributed environment
then suddenly a whole other challenge comes up how do these different micro
services find each other how did they discover what do you do when you how do
you communicate with each other how do you get visibility in which services
talking to which service with what particular load that they’re actually
pushing there and also what do you do when failures happen
how can you out traffic away how can you if things are starting to bring out from
a performance point of view how can you float all your clients all these kind of
steps that you need to take if you suddenly live in this complete
distributed environment so for that we built a delirious app mash that makes
use of the end for sidecar capabilities that actually now give you a complete
view of the network where you can have one consistent mechanism between the
communication for all of the different components that now live in your
distributed system and actually need scare of the reliability of the
communication of failure isolation and actually gives you also insight into how
communication is happening about particular loads in what particular
paths are being created and also how to configure this so all these capabilities
in app mash I’m happy to tell you today they are also general available for
everyone to use yeah and so this whole move to micro services is is is very
important and it is receipt is happening not only in young businesses but
definitely also a more established enterprises so every May is a financial
services company technology company that is moving all along to AWS and they’re
really embracing service to her more about the move to the cloud which
welcome Satish hvala the SPS cloud engineering of le me on
stage thank you very much good morning
everyone I’m really excited to be here to share a
Leamas journey to the cloud who is Ellie Mae
sorry slides for a little bit late so Ellie Mae is a technology that powers
American dream Ellie Mae’s mission is to automate everything that is automatable
in the mortgage industry to make home process easier for home buying process
easier for lenders and home buyers because at the end of the day home
buyers don’t dream about a mortgage they dream about a home today 40 percent of
all US mortgage –es are processed using elements technology in today’s reality
not easy or efficient for lenders or homebuyers so it’s complicated and
disjointed process with dizzying number of steps so there’s a better way Alima has built a platform that helps
solve this problem for our lenders to originate loans more efficiently and
make it make better decisions based on the data let’s take a closer look a robust developer community Ellie Mae has
elements lending platform is a two-sided platform with lenders are one side and
home under consumers and borrowers on the other side consumers and lenders and
partners use this platform to process mortgages every day
we have a community of $5,000 purrs innovating on our platform every single
day our journey has began in 1990s with a client-server architecture built for
on-premise then we transition to SAS in 2009 then we transformed ourselves to be
a platform company in 2016 and that’s built on AWS Alima is moving all in to
AWS our goal is to move 100% by end of 2020 this many benefit there are many
benefits of moving to the cloud given the seasonal nature of our business you
probably know most homebuyers buy homes during springtime or summertime so
elasticity is key for our business in addition to that developer productivity
and speed of innovation is key as well let me give you an example of one of the
new products that we launched we built an end-to-end data pipeline and data
products that takes every data transaction and stores it in a data Lake
built on AWS provides analytics and insights on loan activity for our
customers we built this Pat product from idea to go live within six months this
could have taken 2x longer if we had to build this on on-premise in one mind
let’s take a look at some of the AWS services
Alima is using so like most of you we leverage wide variety of services I
think one or touchdown whew we love lambda it not only saves money it
increases developer productivity significantly
the fact that we brought the matter of fact we processed 1 billion transactions
in the month of January last month alone out of that trillion transactions that
he was talking about speaking of saving money we did some cost analysis in phase
one we are projecting twenty percent of cost savings when we move all in on
three WS and more in order to ensure you know as we continue our transformation
we are anticipating much more deeper savings as we transform in order to
ensure success the key for us is to enable a cloud culture within the
organization moving the cloud in a regulator regulated industry like us
there are a few things we need to consider we need to ensure that our
compliance and security requirements were met while ensuring all the internal
stakeholders are aligned to help kick-start accelerate our transformation
we engaged in a number of cloud centric activities I’d like to share some of
them with you some of the key programs we have implemented included bootcamps
cloud gameday hackathons and technology summit AWS is a key partner with game
day and meters as you all know great people make great
products a happy developers build amazing products I’m proud to say that’s
what our team does our eleme every day thank you so if all of this these new
patterns that we see arriving is not just the architectural panels of course
that you have to keep in mind you also have to think about so there was the
operational model how I’m gonna operate my services and and that the dog feels a
little bit leaf sort of whether you choose containers or instances or lambda
to build your applications into because you know there is a clear increase in
complexity when you move to such a pervasive microservices environment if
you look at it just already talked about a three moving into versed well over two
hundred three five different micro services but if you look at some of our
customers they’re easily winning thousands of different micro services as
part of their overall system and so is that was it easier in the days we ever
in everything in the moment yeah for some parts definitely things were just a
function call or procedure call now you have to use app mash to sort of stitch
all these pieces together I managed the reliability and the fault isolation
there but there is all these different choices you have to make and so what’s
the best operational model around it such that you sort of minimize minimize
the return that you have in terms of how to stitch these different services to
together and you know whether you pick server full and I consider instances as
well as container services that are winning not over phytate I’ll consider
them to be server for yeah because she still have to manage
the underlying infrastructure for that and containers of a far gate as well as
using London in terms of the compute side of it is is I think sort of the
first choice today and we see most of our customers actually really embracing
service as a cloud first strategy except for maybe sometimes you have you know
pre-built software that you came from a vendor and you still may need to run
that in any instance but you see most of our customers building then starting to
build things around it using long that and and service capabilities but it’s so
much more service by the way then just lambda lambda was just the last piece
that was needed to stitch things together such that you never had to
think about service anymore now I think the general model for service is really
that you have no infrastructure to provision it scales automatically you
only pay for what you’ve used and the service itself manages high availability
and security for you and that’s not just lambda I mean it’s all the different
capabilities that we’ve seen at AWS over the years s3 is service matches this
description perfectly DynamoDB now or a world as service or
all the other integration capabilities that we have to stitch your applications
together whether they’re the step function so SNS and SQS and api gateway
and app sync and then as computational models lambda and forget surfer this is
a whole stack it’s not just compute and just focusing on functions as a service
is not service it’s the whole stack in all norm of these pieces you have to
worry about the proper provisioning you don’t have to worry about sort of
military deployments it is all taken care of for you under the conference and
that’s truly what services and it really helps many of our large customers moving
significant pieces of the infrastructure over into a surface and
well financial engine saved 95% in deployment and operational costs
yeah coca-cola cut some processing time from
36 hours to 10 seconds and FINRA the government organization that monitors
the stock exchanges for fraudulent and anomalies and anomalous operations
literally validates about 500 billion stock market transactions a day using a
service environment and so it is really the first say the cloud first strategy
to look at service if he can build it there because they no longer have to
worry about your infrastructure that you need amendment so and I’ve said this
before surface really pushes it out to the limits where in the future you
really will only write business logic nobody will be managing infrastructure
anymore what you operate is a higher level constructs of micro services now
if all of that of course the way that we build software and deliver software
needs to change at all and if you often look at sort of the questions that we
get asked when we think about micro services development is yeah so how does
the release process work how do you push code out how do you debug it how do you
because all of this is such a new environment that all the tools we have
actually need to adapt to them as well and so in the in the old life cycle
things are clear now yet one pipeline that actually delivered into production
maybe every three months or maybe every six months depends a bit on what kind of
development strategy were using but most companies that are actually running a
model if really running sort of a waterfall models as well of course the
great thing with all those micro services that the development life cycle
is very differently each of these teams are fully independent and that means
that they were able to really every team can deliver in there at their own pace
and immediate reacts to requests from from customers
in the old days it was much harder if you have a monolith to actually be
really agile and fast-moving and immediately react to your customers
because basically all your teams are working on the same piece of software
and that has a very heavy weight development process and so best
practices around all of this is still really try and not only decompose your
architecture into smaller building blocks but your organization as well so
they can actually move fast that each of these teams have total ownership over
the software that they have and can actually really move fast based on the
feedback that they are getting from their customers and all of that
infrastructure is code and automation and all this kind of things play they
play crucial walls now these are all best practices that we’ve seen arrive
over time and many of us have all have our favorites kind of the development
tools and many of our partners are the delivering great technologies there and
of course in AWS you needed to make sure that we have AWS cloud native tools as
well yeah and so the whole code pipeline we’ve called called committing deploy
and sort of all the different testing tools that are available around it and
also integration with x-ray in cloud words you need to make sure that we have
at least as a whole set of very mature development tools for you to take you
can automate these pipelines and especially now with the rise of service
at the rise of lunga we need to make sure that all of our development tools
are really supporting these containers and lambda as well and so and most
important of course in all if you build your systems you need to be able to
debug them or at minimum you need to get a good idea in into them and so AWS
x-ray allows you to get a visualization of all the different components of your
micro services environment and where those are waiting in containers or where
they’re running in in lambda and really whether or not you use app mash there
it’s integrated in all of that and the switch you can get a visualization of
any challenges or any problems that are happening in your completely distributed
environment right and it definitely when we think about sort of resource
provisioning you know you may have a case where you said certain read and
write capabilities or you DynamoDB instance on one hand and five micro
services down the path is actually affected by how you have proficient that
now in a disability environment that’s pretty hard to figure out exactly where
these challenges are x-ray gives you detailed inside and view of how you’re
distributed systems application works now with all of that debugging is
important it’s a very important part of our development process so with the rise
of surveillance we need to make sure that you can use your most popular tools
to actually really build service applications on AWS yeah and we had
already announced cloud 9 of course which is our which is the AWS ID
environment that is truly cloud native but also pycharm and actually today the
IntelliJ toolkit is general available so you can build your java and python ones
free s code is still in Developer Preview but we expect that to go general
available as well soon and so for all of these and we and know we’re pretty
opinionated about what are the best development tools we have to use yeah we
just need to make sure that all of them work really well on AWS and that you can
develop in the way that that really is your style of development and that goes
from programming languages all the way over to what kind of IDE you want to use
although I don’t see any integration in VI and Emacs happening anytime soon
yeah but we do look at all the different kind of models and so for example Sam is
is a different way of actually describing your service application in a
way more declarative format instead of imperative and so check out Sam if you
really want to do local development as well and it really helps you sort of
think differently about how to pose your server this application now in
all of that I think there is wrong really crucial point that we all need to
start thinking about as technologists you know and that is that in this whole
of continuous integration in the development world if deployment world
security certainly becomes very different and and I think that we as
technologists really need to take responsibility for making sure that we
keep our customers and our businesses secure even if we changing operational
and architectural models as fast as that we’re doing now if I look at some of the
past and probably sort of the monolith idea you know you would build software
the security team would come in they sprinkle some magic dust over it and
suddenly your application is secure well I think that may have worked maybe in
the past but I think today that’s definitely no longer the case I think
mists of these old-style security approaches all rely on the fact of
building firewalls around it yeah now if firewalls were the right security
solution we would still have moats around our cities we don’t we protect
our individual houses we protect our individual rooms within our houses so we
should do that in our digital systems as well I remember most if you look at most
of the threat data or the security data it shows that this brute force from door
attacks almost never happened anymore it’s all about social engineering where
someone in your organization will get an email that says oh this is your new
retirement package click this link to sign it there is always an idiot that
clicks that link because if not they wouldn’t be doing it with it yeah and so
there’s always some some Evo JavaScript can download that you know gets
established and if they’re not the individual pieces in your organization’s
are individually protected everything is toast and I think what we see in the
past years with the number of data breaches
have happened most of those or almost all of them are related to sort of old
digital systems that have been brought online or old operational practices that
were appropriate maybe five or ten years ago where you were building it according
to a waterfall model and things like that those are no longer applicable now
that the security team looks very different today than 10 or 15 years ago
now we have all these different components that we have to take care of
and it’s no longer the secure separate security team it is us as builders that
are responsible for this security needs to become everyone’s job and with all
these data breaches that we’ve seen in the past years we need to make sure that
we protect our customers and our business and it’s our responsibility as
technologists we really need to make sure that now that we are moving to more
and more digital systems and most of these digital systems are developed in a
very different much faster moving way that we do not forget that security
needs change with this and it’s both a security off the pipeline itself as well
as the sort of software that you develop inside the pipeline and make sure that
your pipeline has high rate development services that you have total control of
and then make sure that all the components that you’re building actually
in each step of this that you check against our we make our introducing new
vulnerabilities or not what kind of alarms should go off if you do a hundred
deployments a day security is look different so if this is sort of a
traditional set up for your continued integration and deployment we need to
make sure that in each of these steps that are happening Security’s become
impairment yeah and so whether you do the continuous scans whether if the
changes happening in configuration alarm bells need to go off and sometimes
either automated or sometimes manual checks need to happen if someone adds a
new library to your application but no Lyle needs to go off because someone
needs to see whether this is actually the library that that’s been approved
that maybe it’s an open source library whether the vulnerabilities that come
with it all these kind of steps and why is this library being added and then all
of these kind of things you need to make sure that you can automate that as much
as possible and if you look at sort of the different components they’re
definitely infrastructure has codes place it plays a crucial role in all of
that because that means that you can actually see the changes that are
happening in your infrastructure configuration between one in the other
and so if we look at when in this whole development process you need to actually
have all sort of operations happening it is both we push new code that actually
needs to go through this go through code scanners it needs to look at sort of new
libraries and your dependencies being integers or you know whenever this event
has happened and if and triggers can be you know whenever a change happens or
whenever you maybe you need to do it on a daily basis or whenever you change
frameworks and then afterwards you need to continue to validate validate whether
actually the application is still meeting your security and maybe
compliance requirements in all of that automation plays a crucial role and of
course in the sort of in the world that we live in there’s this shared
responsibility model a DBS takes care of a large part of the operational
environment for you and you build also all these new tools for you to use
there’s a whole collection of Adria’s automation tools around security that
you all should be using because I think that really if you really want to move
to a world that is secure you need to automate as much of the security
processes as possible now let me just pick a few of these that I really like
so Amazon inspector basically contain inspect your code that you’re running
whether you have introduced new film abilities in one hand that might be just
skinny against well known vulnerabilities but it might also be the
case that you are subject to particular compliance regulations an inspector can
actually check whether you are still in compliance and this is important we
remember that you maybe need to process credit card transactions you need to be
PCI compliant and if you now make a home that changes to your code a day are you
still in compliance in compliance with the regulations yeah an inspector can
help you with that by really diving deep on some of the changes that you’ve made
in a completely automated fashion so make use of it climb trail if you’re not
being not enabled cloud trail then you’re really missing out on getting
detailed information about how your systems are being being used cloud trail
really locks every possible operation on every object on every resource that is
happening yeah you can continuously record all these API calls and then
launches them into stores them into a sea over which you can have client warts
and other analytics tools then then run over and take a look at sort of really
other anomalies happening in all of this you think about security as I said
before encryption is what it’s all about yeah so dense like no one’s watching and
encrypt like everyone is yeah import it in all of this is that encryption is the
tool we have to make sure that nobody else has access to our data yeah so for
example with Amazon the retailer in DT PCI compliant that means that about 15%
of your calls and storage operations need to be encrypted just decided to
encrypt everything that means that none of your engineers can no longer make a
mistake about should they encrypt this Chile not encrypt it and your PCI order
it’s becoming really simple in that manner so given that we’ve
built encryption into almost all of the AWS services make use of it you know
five ten years ago we may have had this discussion whether HTTP was too
expensive now every every consumer service runs over HTTPS the same goes
for encryption for a long time you sat down how these tools are way too hard to
use and it costs too much and it turns out we’re now building these tools that
don’t make it that difficult for you and whether you have integration of
encryption into all the different services or you know whether you can
just bring you and keep if kek you get with kim s i remember this means that
you have total control over who is access to data look at redshifts
redshift encrypts every data block with a random key always and then the set of
random keys is encrypted with a master key now you can bring your own master
key or we can generate it for you the most important if you generate your
master key you’re the only one who can decide who is access to your data so
encryption is the most important tool you have to protect your customers with
all of this I think as technologists we need to take responsibility here we need
to make sure that the next generation of systems that we’re building have
security as a first great citizen where you need to start thinking about
protecting your customers on day one and I know if you’re a young business here
you start to innovate is that think about all these new things that you want
to build security might not be on the forefront of your mind but it should be
and definitely us as technologists need to take responsibility for that making
sure that the next generation of systems are as secure as they can be using the
automated tools that we give you well in all of this there’s also changes to data
and data management that are that are happening of course and that we at AWS
needed to make sure that you have the right tools to use and so many of our
customers for running databases themselves enterprise-grade databases on
on premise many of you have asked us to to help you move to open source and
whether there is my sequel or Postgres Reeb mostly not because of the
capabilities necessarily of these database engines but because the
licensing terms that the old guard is using is truly restrictive it’s almost
goes back to blackmail yeah the only way for you to drive your costs down is to
make very long-term commitments and then you know buy many more licenses than you
ever need well I’ve been on the receiving end of that if I had to buy
more databases the only way to drive course down was to make a five or a ten
year commitment I don’t know how many databases Amazon will need in ten years
from now but that was something you need to decide at that moment so many of our
customers really want to move away from that sort of restrictive environment I
really want to move over to the cloud preferable using standard interfaces
like my sequel or Postgres but really really would like an enterprise-grade
database in the background now none of these relational databases actually have
been designed for the cloud the only way that we can really scale them out
instead of scaling up is to actually make use of sharding yeah and whether
you do that at the application level or if they do that sequin level or whether
you do it in some weird storage making mechanism
those are the ways that you can make use of these databases to scale them out but
remember this is technology developed in the 90s it’s not modern development it
all requires a local disk or even if you coaster of databases they still requires
a shared disk and each of those each of those instances I have a whole stack in
them but actually very much duplicating everything so with all of that we did we
built Amazon nawara where we basically we up the whole database engine apart so
Aurora has two interfaces as a my secret interface in the post-crisis interface
but behind the covers we’ve ripped everything apart more or less at the
middle of the caching layer I moved to a shared storage service
based on SSDs which is actually the database aware and this is sort of has
allowed us to build a much faster much higher reliability system than we could
ever build using sort of the standard off-the-shelf databases now we actually
make use of 6xo application so if you build this distributed storage engines
you use quorum technology to sort of make sure that you know you can actually
read the last right and so typical scenario is there where the quorum is
sort of three nodes and you need to have at least two nodes available to write a
you need to have two loads available to read such as there’s an overlap you can
always read the last right now in our scenario we believe that our Felicia now
is out there that are much more dangerous to to the reliability of such
a database in our case we would really like to survive at least the failure of
one complete AC and wow there’s a complete I see that I may have failed in
that particular timeframe there’s a likelihood that one of your other nodes
may fail as well it’s just you know when it when it rains it pours and so we
really want to make sure that we have an AC plus one failure scenario where we
can lose a whole AC and one note in the different education either go down to a
quorum system of six so we do six way replication to make sure that we have a
continuous overlap in these scenarios so that means that if you lose an AC a
loser note you may no longer be able to write but you can we still read and then
we need to make sure that the repair for writes is really fast and we do that by
making sure that the individual blocks that are being stored in the storage
service are really small so it’s ten gigs meaning that you can very quickly
we repair a field right but actually we replicating the data underneath them now
in all of this that’s the reliability side of things it is also very important
to make major improvements so – the way the performance in this
traditional databases is restrictive because their whole thinking is about a
local disk if you look at a typical my sequel writes is actually any write to
the database will result in five writes to the storage engine which then writes
five times to the backup engine they write all the data to your replicas who
then also does all of these storage operations yeah this is hugely expensive
and so you remove the data pages you do double writes to avoid corruption
happening you move the logs you move metadata all these different pieces are
being written now it turns out that that’s hugely wasteful it turns out that
it’s only the log that you actually needs to write because in the log you’ll
find it before in the after picture of your database and so we don’t need to
move to data page now you can just moving the lock and so in the world we
only move the log write a log to the storage changes and then the storage
engines are not just storage engines the database aware because you can actually
recreate a database by just purely looking at the log and the only reason
why you would ever need to move a database from your story changing into
the database is if there’s a cache miss well it’s most likely that the most
recent transaction that you’re actually completed will still be hot in your own
cache so you can actually recreate these data pages in a very lazy fashion what
you see here is that the primary instance right twice the log gets
persistent at that moment you can delay the acknowledgement event right then
gossips with the other storage nodes your six storage nodes to actually
transport data there and then in a lazy manner you can start recreating your
data pages all of this has allowed us to create a foundation for to innovation in
databases that was never before no other database systems can do this kind of
innovation because they’re still stuck in the old architecture decompose it
into smaller building blocks and then apply standard distributed systems
techniques to to actually keep him reliable and
performance and so all this gives us a basis for database innovation that it’s
pretty spectacular if you firm if you ever programmed in a
language that has the object relational mapping available that grew beyond Wales
for example any changes to your data structure will immediately result in a
change of your schema for that all cell databases need to do complete table copy
however in Aurora basically copying creating a new database or creating a
new table based on the old table is a matter of microseconds because we can
recreate we can create a new table in a lazy manner because the storage is data
big blog aware so it’s been a great success fastest-growing service in the
history of AWS and it’s still growing very fast mostly also because we were
able to push all these new innovations further and it’s not just relational and
we talked about this earlier definitely the move to micro services has made
everybody aware that how wait maybe I can pick the right tool for the job this
particular micro service just needs a graph database yeah or maybe you’re
operating in the world where you’ve been considering blocks change style of
interactions were you looking for an immutable letter they make use of key
ODB so each and every one of these services serve a really particular
pattern dynamodb has its roots in in a deep dive that we did and Amazon the
retailer itself did it in 2004 and when we did a deep dive on how we were using
relational databases it turned out that 70% of the uses of these relational
databases was key value there would only be a single key in the query you would
get a single well back 70% what we knew that we could build very different types
of databases that will be uniquely positioned to serve a key value world
and we could have title trove of performance reliability all of
those and dynamo became that and DynamoDB later became the service
version of that that we had an AWS but again it turns out that DynamoDB is a
powerhouse now for everyone that wants to do 2d scalable operations if you
think about supercell for some head the company that makes Kings they made these
games clash of clans and others on day one of
a game they will literally have millions of players checking out the game that
means that your data stores behind it need to be extremely scalable because a
bad experience on day one will not have gamers come back and so dynamodb is the
powerhouse that sits behind all of that now everyone is looking to get more
value out of their their data I mean one of the things that client has done has
made me the whole ante landscape egalitarian everybody is access to the
same storage the same compute the same databases the same analytics tools the
same IOT tools the same ingestion tools everyone has access to that now so IT
capabilities are no longer competitive different shield so what is them the
differentiator it’s the kind of data that you have and how smartly you make
use of that data and so we need to make sure that we can actually help you pick
exactly the analytics tools you need to use to operate on your data and whether
that is sort of in the analytic space or how to create data leaks or how to move
data in and out of out of your data Lake yeah and so all of this is crucial for
you to pick exactly the right tool you want to do at work preemies lucifina you
want to make use of Hadoop You Xiomara you want to do this do very complex
traditional data warehouse create queries you make use of riches and so
pick exactly that right tool for the job well after all that shift is a data
warehouse so you can just fire up on the mountain where in the past maybe data
warehouses were something that was very expensive and centralized you all needed
to cure for that what we now see is that many
business units are just firing up a data warehouse for two hours on the first day
afternoon that is a radical shift in how databases and data warehouses are being
used including at Amazon and I’ve trained before that of the past year
November 1 was one of my happiest days of the of the year when we shut down one
of the world’s largest it’s not D largest or called data warehouse and we
placed it and we placed it with red shifts at Amazon and so if all of this
we have truly moved to an environment that moves so much faster so much more
agile because indeed in the old-style data warehouses it’s such an expensive
piece of software and hardware that everyone these are loaded up to the max
you always need to wait for it especially if you want to run some ad
hoc queries try and forget that you’ll go back in the queue where maybe your
queries get executed tomorrow and it’s really absolutely becoming the most
popular cloud data warehouse out there because it’s so easy to instantiate and
so into it moved all of the mission-critical
analytics workload over to to redshift instead of it’s on premise environment
and so much moving so much faster the cool thing with richest is that we’ve
enabled deep metrics application and database metrics into the into the
system and as such we’re able to really observe how our customers are using our
software and then working with our customers understanding how we can
actually speed things up for them in the past two years we’ve been able with all
these improvements to speed to make redshift 10 times faster mostly because
really this close interaction with our customers really trying to understand
what are the kind of things we can do short query acceleration or resizing
elastic resizing or how to speed up interactive queries all of these kind of
things is working together with our customers understanding
the patrons you use in a model in data warehouse and that as leopards enormous
speed improvements over time based on the feedback of our customers now all of
the other things I just talked about is sort of waiting for your cribs now turns
out that there is it 87 percent of our customers never wait for their queries
but once the other 13 percent what are the kind of things that we can do in
terms of innovation in our data warehouse to make sure that you never
have to wait so if that we’ve launched redshift concurrency scaling which is by
the way today generally available so what does concurrence is scaling do we
basically make blur said burst clusters available for you that if we see the
queues queues of queries sort of rising to the point where you actually have to
wait to execute it clearly we can fire up additional additional clusters for
you such that your customers never have to wait and so much of this comes
actually at no cost at all to our customers yeah because we will actually
sort of fire up these queries for you in this cluster without actually charging
you extra for that now analytics plays a very crucial role and
we have to think about analytics us oh yeah that’s sort of the data warehousing
this the old-style world but if you look at some of every modern young business
or every modern young application is being billed as data generation and then
the lytx integrated into it well we all know about fortnight nobody here plays
for tonight yeah wires oh by the way I serve I’ve got more than five minutes so
I can claim I’ve done that more importantly next to sort of all the
efforts that the epic guys have put into building for tonight as a game they put
enormous amount of effort into data generation around that and really as the
game clients for the servers or different types of pieces that all
generate data for them there’s a massive analytics environment sitting underneath
that serving the pieces of the business or wall-mount
real time like service health and tournaments yeah but on the other hand
also just business capabilities like just measuring your KPIs or actually
analyzing how the game is being used such that the next generation of your
game that you’re building is actually meeting sort of the ways that your
customers are playing it and so I’ve always looked at sort of in these things
as analytics having three different pillars one of them is looking backwards
yeah I’m looking backwards really means sort of this the redshift type of
operations EMR where basically you basically generating reports and then
there is the real-time part there’s a real-time pillar where you use Kinesis
and elastic search and EMR to probably look at what is my inventory level right
now I’m not interested in an inventory level yesterday I want to know what it
is now and that is real-time operations and then there is different one yeah the
third one is how to predict the future and so looking backwards what’s now
what’s the future now we’re really bad I think at predicting the future so the
best next thing that we can do is make use of data that we already have and be
smarter with it using AI and machine learning so with that I’d like to invite
dr. Matt bouddhiste general manager of deep learning in artificial intelligence
to talk to you more about that Matt’s good morning everybody and thank you
Verna so as I’m sure many of you are aware we’re entering a new golden age
for machine learning where many of the constraints which have held back the
application of artificial intelligence and machine learning to real-world
problems start to melt away in the cloud and as a result of that we’re starting
to see tens of thousands of companies in virtually every industry and eventually
every size and shape start to apply machine learning to their central core
challenges whether it is change in health care through change healthcare
whether it is advancement in life sciences with folks like bristol-myers
Squibb and Celgene folks progressing manufacturing allowing you to operate
more efficiently telephony contact centers you name it machine learning has
arrived in virtually every industry and it’s incredibly exciting to be part of
the team at AWS which is helping customers drive this forward and a big
part of why we’re seeing this stratospheric movement an advancement in
machine learning is that on AWS there are a number of really key tailwind
these are forces and services and capabilities which are available to
developers just like you which drives significant acceleration in your use of
machine learning and what I’d like to do today is just run through the four key
tailwind that we’re seeing in the trenches at AWS and run through what I
think are the key challenges and the key solutions which are only available to
customers today on AWS the first tailwind which is driving developers to
do more with their data is a broad and deep set of capabilities which aim to
put machine learning in the hands of every developer we joke internally that
we would just want to make machine learning boring we wanted to do just
another tool in the tool chest which is available whenever and however you need
it and to do this we make three main areas of investment the first is an
investment in the fundamental machine learning frameworks and infrastructure
related to machine learning so this is typically where the advanced machine
applied scientists live as they’re building our advanced models researching
new ones or even iterating on the key frameworks themselves and these
frameworks are how you define your neural networks and your workflows to
train your models and then that’s where you run the inference to make
predictions against your models they’re almost all open source they have some
strange names such as tensorflow MX net and pi torch there are other high-level
interfaces such as gluon and caris and our approach here is maybe a little bit
different from others our approach here is that we want to support all of these
incredibly well and make sure that they run as well as possible up on AWS and
the reason for this is that as the science of machine learning is advancing
new techniques and models and approaches and architectures are being made
available virtually every single week and those architectures they exist in
all of these different frameworks they’re published with reference
architectures in all of these different frameworks and so just picking or trying
to standardize on one is not the right approach because you lose access to all
the other innovation which is happening in all the other frameworks so our
approach is to invest in all of these areas and we actually have separable
teams on at AWS which focus on tensorflow and MX net and PI torch and
so on and we’ll keep doing that as more and
more these frameworks start to appear so part of their approach is we want this
to be as easy as possible for developers to use and so we take all of these
frameworks and we run them on world-class infrastructure that a lot of
you are familiar with on ec2 and we make it available in different ways we make
it available in a fully managed service which we call sage maker which I’ll talk
a little bit more about in a second but we also make it available in an ami or
an army where we take and optimize all of these frameworks and just make it as
single click to deploy them up on ec2 and this DIY approach is really popular
with scientists and apply machine learning developers who want to get in
and tinker at a very very low level and potentially even build more frameworks
going forwards but as Vern has been talking about we see a definite trend
with more and more developers turning to use containers and so we want to apply
the same approach where we’re packaging optimizing configuring installing all of
these frameworks and make them available not just in an army but as a container
and so today I’m very proud to announce AWS deep learning containers these deep
learning containers allow you to quickly set up deep learning environments up on
ec2 using docker containers they run on kubernetes or ECS and eks we’ve done all
the hard work of building compiling generating configuring optimizing all of
these frameworks so you don’t have to and that just means that you do less of
the undifferentiated heavy lifting of installing these very very complicated
frameworks and then maintaining them because they all move very very quickly
and we’ll be releasing new containers as new major versions are made available
for tensorflow and MX net and we’ll be adding PI torch
very very soon they’re available in the AWS marketplace and through the ec2
container registry so moving up a tier the second major area where we’re making
investments is the machine learning services and our big investment here is
a service called sage maker and what stage maker attempts to do is it
attempts to bring machine learning and put it in the hands of any developer
irrespective of the skill level that they have as it relates to machine
learning and it’s sometimes easy to forget just how challenging machine
learning used to be before with the introduction of sage maker virtually
every step of the machine learning workflow presented a hurdle or a wall
for most developers who didn’t have deep skills in machine learning or deep
learning and combined these walls were effectively infinitely wide and
infinitely high they were just if impossible for most developers to climb
over or dig around but with sage maker we systematically approached each of
these key challenges and started to remove them behind a managed service
which I which is very very easy to use so for developers who need to collect
and prepare training data this is everybody by the way that wants to do
machine learning pretty much we replace that with pre-built notebooks for common
problems and a managed notebook service which with a single click gives you
a notebook environment where you can start to experiment and slice and dice
your data instead of having to choose and optimize your own machine learning
algorithms we built in a set of over a dozen high-performance algorithms these
are optimized for AWS and we use some clever techniques to allow them to
stream data from s3 and train in a single pass which dramatically increases
the accuracy you can obtain and reduces the cost of running them we allow
one-click training so with a single click we can spin up a fully managed
distributed cluster under the hood for you to run your training against and
then we added optimization so a dirty secret of successful machine learning is
that you don’t just train one model you train a thousand and just pick the best
one and this has traditionally been kind of a trial and error approach but
instead of that in Sage Maker we have a service which provides hyper parameter
optimization and with a single click we’ll drive and actually guide the
search for the best possible model using machine learning under the hood when
you’ve got a model that you love you can make it single click and deploy it in a
fully managed environment and then scale that environment for production use
using auto scaling so you can scale up and scale down and the result of this is
that more than 10,000 developers today are using Amazon Sage Maker to drive
their machine learning workloads and many are standardizing on the platform
as their machine learning central repository of data and analytics related
to ml the third main area we want to provide these sorts of capabilities to
application developers who don’t necessarily have any machine learning
experience and so here we provide a set of AI services which mimic in many cases
some level of human cognition and so we have a set of services for our vision so
computer vision recognition to do image and video analysis and text rack to
automatically extract data from scanned documents we do a lot of work around
speech the both the generation of speech using a service called Pali that’s the
same service that we use to generate the voice of Alexa and transcription where
we take speech and turn it into text and then investments in language models
where without any machine learning expertise you can start to apply natural
language processing and translation to the text
that you potentially captured through speech we build conversational
interfaces using Lex that’s the same natural language understanding system
that we use under the hood with Alexa for building conversational interfaces
and just in December last year we announced two new services for
forecasting and for recommendations and this allows you to build our very very
accurate great learn deep learning driven forecasts and deep learning
driven recommendations based on the same technology that we use on the retail
side of the house at amazon.com and what’s interesting about these last two
services forecast and personalized is that unlike some of these other deep
learning systems unfortunately there is no master algorithm for driving the very
best forecast there is no master algorithm for driving the very best
personalization experience whether it is order predicting the news articles or
ordering search results and as a result what you need to do is you want to be
able to take the data that you already have and then train your own models
which are specifically for your data and for your customers that’s by far the
best way of approaching it but the challenge is that this is incredibly
complicated and so what we do here is we apply a technique that some people call
Auto ml where we take in a lot of input data so a real-time activity stream of
what’s going on on the platform in terms of personalized the inventory so the
articles or the products that you have along with any demographic information
optionally that you want to provide to to drive the personalization engine and
then with a single click just three API calls you can build a customized version
a customized version just for you for personalization and recommendation which
we host behind an API on your behalf now we don’t use any of the data it’s
customized for you we don’t share it in any way this is just a specific private
model for your use but under the hood we’re doing a world of things to make
this possible and one of the great things that keeps me skipping into work
every morning is the opportunity to invent and simplify relating to machine
learning on behalf of our customers I think this is an excellent example where
personalized under the hood is using machine learning to make all of these
decisions and we train those machine learning models based on knowledge that
we’ve had during several dozen personalization
systems at amazon.com and then we drive the workflow from loading the data
inspecting the data selecting the right algorithms training the models
optimizing them all of that all the way through to hosting them building the
feature stores and the caches on your behalf so that you don’t have to worry
about it so this is a step function change in the speed at which you can
start to introduce deep learning and machine learning into your
organization’s the second tailwind is that customers are able to take
advantage of AWS to increase the performance of their machine learning
applications whilst also lowering costs normally you have to choose between the
two but we think that’s a false choice and so it’s never been cheaper or easier
to run your machine learning workloads on AWS so machine learning you take the
data that you have usually stored in s3r you run it through a training system and
then you use inference to make predictions and usually do that as I say
inside these frameworks so I’ll just use tensorflow as an example so tend to flow
very popular great tool about 85% of all tensorflow workloads out there today run
on AWS and we see this across virtually every industry whether it is into it or
siemens or startups a huddle they’re all using tensorflow
up on AWS now the challenge with tensorflow is whilst it’s a great tool
it’s got a lot of opportunity for developer productivity once you start to
get to production he trained on very very large amounts of data you start to
get a scaling hit so it’s not particularly efficient when it comes to
scaling across dozens or even hundreds of GPUs and so what we did with our
tensorflow team is we went super deep into the the central engine of
tensorflow and we optimized the networking to be
less chatty and to be more efficient across the AWS network what we saw is
that using our AWS optimized version of tensorflow which is available in the
containers and in the army as well as sage maker you get you can train it
nearly twice the speed so with stock tends to flow you’re operating at about
65 percent efficiency across 256 GPUs that means for every dollar that you
spend only 65 cents of that dollar is used for anything actual good
the rest is just overhead moving to the AWS optimized version we see a 90
percent scaling efficiency now you’ll never get to 100% it’s just it’s just
not possible today but 90 percent is a significant increase in speed and what
that means is you can train your models with more data you can train your models
faster and you make better use of your most expensive resource which is your
data scientists and your developers which are using the machine learning
techniques but another dirty secret of machine learning that’ll let you guys
into is that whilst training is incredibly important there’s a lot of
focus there it’s actually only a fraction of the total cost of a machine
learning workload and running inference in production is the overwhelming
majority of cost when you start to break it down about 90% of the cost of a
significant machine learning system is running predictions against your trained
models and so whilst we’ll continue to optimize on training we’re going to
continue to focus on optimizing this big chunk of work which is in improving
inference costs we’re doing that today with a service called elastic inference
and this allows you to as a service add a slice of GPU acceleration for your
smaller models and then dial up the GPU acceleration at the end of an API when
you need to increase the throughput or you start to work with larger models and
just this service alone can decrease your inference costs which is the
majority by up to 75% you can scale that from a single trillion operation per
second which sounds like a lot but actually isn’t in terms of machine
learning all the way up to very very big beefy models such as up to about 30 to
trillions of operations per second and we’ve already built this into 10 to flow
MX net and will support any model which conforms to the Onix standard coming up
towards the end of the year we’re gonna see AWS inferential start to
be introduced and this is our AWS design custom machine learning inference chip
and this is designed for more sophisticated models those can take
advantage of an entire attire chip with high throughput low latency and with a
single chip operating at hundreds of tops but can be combined together to
operate at thousands of tops we’re going to make these available through ec2
instances through sage maker and also under the hood of elastic inference
so if you start using that service today when we introduce inferential later this
year you’ll start to see it just an automatic increase the third tailwind is
that it’s never been easier faster or cheaper to do to get data ready for
machine learning this is an area where a lot of organizations spend a remarkable
amount of time and honing and producing accurate training sets is one of the
most important ways to build successful machine learning models so these models
require very very large amounts of data tens of millions of images and so if
you’re building say an autonomous driving system what you need to do is
you need to take every photo every frame of the cars that are driving around
collecting this data and you need to annotate it in some way you need to tell
the model through training what is important and what is not important and
the way that’s done today primarily is through humans you show all of those
images to humans several at a time and you get them to say this is sidewalk
this is a car this is a stop sign and these annotations these labels are what
allow the machine learning systems to learn however it’s extremely costly an
incredibly complex to do this at any sort of scale because not only you
managing the data you’re also managing the humans that have to go off and
actually provide the the annotations and so we provide a service uniquely on AWS
called which is built into stage maker which we call ground truth and ground
truth allows you to build highly accurate training datasets which reduced
trainings that data set but preparation costs by up to 70% and we do that under
the hood by using a technique called active learning we take data and as it’s
being annotated by the humans we capture all of that cognitive investment and we
train a machine learning model as we go and it progressively gets better and
better and better and more and more accurate it learns more features as the
training is taking place and this means you as you train more data you can
offload with confidence more and more of the annotations to the system which your
training as you go so with no additional overhead you start to dramatically
reduce as you go the number of images which need to be shown to humans in
addition to that have world-class workflows and tooling
to allow humans to provide those annotations but the key to driving down
cost is to learn as you go capturing that cognitive investment the
final area is that it’s never been easier to learn about machine learning
one of the things I love about the AWS community is just an insatiable desire
to broaden skills and expand their knowledge and on AWS it’s never been
easier to make this investment for yourselves in machine learning we’ve
taken the machine learning University content that we use to train our own
engineers at Amazon and we’ve made that available in a self-service way on
through our training portal this is one of our most successful training programs
to date we also make our own engineers these are folks that have built things
like personalized and forecast the engineering teams that are involved in
the personalization platform over on the retail side of the store and we’ll make
that team available to you to get hands-on keyboards to build initial pocs
so our goal here isn’t to build a big professional services organization we
just want to help spread the knowledge as much as possible through a program we
have called the machine learning solutions lab if you’re more of a
do-it-yourself person as I am then we make some products available to help you
learn one of them is called deep lense it’s the world’s first deep learning
enabled video camera for developers and this allows you to capture data trained
against that data build models in Sage Maker and with a click of a button will
deploy them directly onto the device these models are actually running on the
camera and then pretty much everybody has things on their desk for object
recognition and people that they can use and with this fast feedback loop you can
start to learn an experiment which is a fantastic way of broadening your machine
learning knowledge reinvent our developer conference in Las Vegas last
year we also introduced AWS deep racer this is a fully autonomous 1/18 scale
race car which is driven by a type of machine learning called reinforcement
learning you build your models in a simulator up on the cloud you specify a
scoring function very easy to do without any machine learning knowledge required
and then you use that scoring function in a simulator to train a racing model
which you can deploy down onto a car and then race around a track and when we
started doing this at Amazon we saw very very quickly and we should
have seen this coming our engineers started to race these
devices and so we’re also announcing the AWS deep racer League this is a global
racing league that anyone can top participate in you can build your
reinforcement learning models up in the cloud we’re starting a series of deep
racing league races at the AWS summits across the world I encourage you to
attend them all there is credit for doing more than one and the winner from
every single race at every single summit the person that has the fastest time
around our test track will win an all-expenses-paid trip to reinvent to
participate in our championship cup in 2019 we’re also running if you can’t get
to a summit or you don’t have a car a series of virtual tournaments running
every month through the year so I’m very pleased to announce that this is
starting today you can head down to the expo you can take some models you can
start racing them around the track we have a real professional commentator
from spoke motor racing generate the the tracks it’s a lot of fun and we have a
leaderboard that you can all look up on your phones and track how you’re doing
so across all of these services the capabilities made available on AWS are
remarkable they are more broad and more deep than anywhere else and all of these
tail winds across price-performance across data preparation and of course
these learning capabilities are only available on AWS and they’re
specifically designed to help developers and builders like you get up and running
with machine learning and to tell us a little bit more about what they’ve done
a word day I’m very proud to introduce Ellen who is the head of data science
and architecture thanks very much thanks Matt good morning everyone would
you believe that we spend close to 2,000 hours at work every year sometimes it
feels like more than that what I love about work day is that vataj 37 million
lives and make their 2,000 ars-art work better and brighter food is the leading
provider of enterprise cloud applications we deliver applications for
financial management human resource management analytics and planning Berkeley delivers an incredible trusted
system of record for some of the largest companies in the world we serve much of
the fortune 500 my background is in machine learning and is in building
machine learning and data products and I’m passionate about using machine
learning to solve some of the hardest problems in enterprise software on top
of the system of record the incredible trusted system of record we have we have
a layer of engagement that delivers reporting on all tags and planning my
team in workday is focused on delivering a system of inside using machine
learning that helps our customers do their best work we have identified a few
areas their mission learning makes a big difference for our customers we all know
how hard it how hard unimportant it is to hire and retain the best talent we are on a mission to transform how you
identify hire and retain your best talent in the world of financials having
the right inside at the right time and right contacts is everything we are
transforming the financial systems which streamline workflows on powerful
predictions today we all expect both wise experiences in every aspect of our
life we are focused on delivering personalized recommendations that make
you better inform more productive maybe even a little bit smarter to do all
these things we needed a solid set of tools on the right partner to get us
there and get us there fast at workday privacy and security of our customers
comes first so naturally first we needed a solid
foundation that ensures the privacy and security of our customers and a system
that enabled us to track data lineage at a fine granularity so that we can
implement privacy by design once this system was in place the foundation was
in place we gave our data sign pairs the best
machine-learning on data tools and of course the fastest compute for them to
train the machine learning models we selected AWS as our partner in this
journey and built our machine learning environment on AWS using a variety of
services this diagram illustrates our AML workflow and the services that we
are using let me use an example to illustrate how this words
one of the financials product we have is mobile expenses imagine you are on a
business travel and you have a ton of receipts that you are collecting as you
expense for a variety of stuff mobile expenses allows you to take a
picture of your receipt and file your expenses on the go under the hood we use
sophisticated deep learning models to extract the details from your receipts
and populate an expense report for you the receipt stem cells are stored in the
data lake and our data scientists use sage maker an MX nerd to train deep
learning models on GPUs once the models are trained we deploy them as restful
Web Services in our data centers we are excited about the potential of using
ground truth now to label these receipts without having to leave our data centers
or without having sorry without having to leave our secure ml environment on
top of AWS data scientists and engineers love the AWS tools and as you can
imagine when you have data scientist and engineers being happy about the tools
they are you saying this is resulted in increased productivity and more
importantly fast experimentation by leveraging sage maker algorithms on GPUs
we have reduced the ml development time from months two weeks my team is like
really excited I’m moving really fast we are looking at a few other AWS services
particularly around the machine learning services that are very interesting as I
mentioned earlier we are evaluating ground truth to label data for our
machine learning there are other AWS services and sage maker features
that are of interest to us like the elastic inference and for some of the
future use cases we are looking at higher level AWS services like our AI
services like Amazon recognition when it looked to ver be there about a year ago
we had a small team of data scientists and engineers with big ideas to
transform the enterprise software using AWS and it’s full suite of services we
built a secure and robust platform and on top of that we are building and
delivering machine learning features to our customers
it has been an incredible journey and we have only just begun thank you section great customer stories today if I want
to do a little bit of a recap here so if I think about sort of model and
application development there’s a number of areas that you really need to pay
attention to yeah first of all really thinking about sort of what are the new
application architectural patterns like you know service first and really modern
applications are truly service and we see really companies making the jump all
the way from mainframe immediate company leap frogging all the way over to
service really pay attention to what kind of data is generated to both help
your operational performance as well as your business performance and what kind
of information can be tree from that to build your next generation of products
and with all of that I really want to emphasize that security is everyone’s
job now because in the future it will it’s us as technologists that will need
to be that are responsible for protecting our customers and our
business in that sense now we’re very fortunate over the past years to meet
many of our extremely exciting customers where those are young businesses or
established enterprises that are going in completely new directions very
fortunate to meet with them and one of the things that we’ve decided to do is
to make a to make a TV series out of it yeah and so we have this long-form video
contents called now go build but basically I visit young businesses
around the world and do a deep dive on how these companies are actually truly
changing the world around them and so the first one that we launched during
the event was this a company from Jakarta called Hara talkin speaking use
of blockchain technologies to to build identities for this poorest farmers in
Indonesia such that they no longer need to go to a loan sharks which will charge
them twenty to sixty percent on their small loans but actually can really go
to a bank because now they do have an identity
I’m actually not only an identity they have information about sort of the plot
of land that they have they yield to the growth and things like that really build
opening up the world of sort of government assistance and things like
that for these youngest farmers it’s a great story
if you haven’t seen that one yet please go see it because these guys are really
changing the world for the poorest farmers in the in the world now today
we’re actually releasing the second episode where we go to you to Singapore
to a company called simplistic we made something called multimatic sort of
really changing the world that young Indian women have to live in by not
having to continuously make food for their families who spend an hour and a
half a day making wealthy stew to eat with sort of really changing the world
and so they they sold 40,000 of these machines have a dubious IOT integrated
into it and machine learning but basically they have a machine learning
driven roti maker and new episodes from Norway and Germany and South Africa and
Brazil will be released throughout the office now this one the next one that’s
coming up let’s take a look at the trailer for for Singapore our planet and
our civilization are changing faster than ever
this is Malcolm will join me as I travel the globe talking to startup founders
using technologies to make our worlds more interesting accessible and livable
these are the interpreters that that Canadian the future so yeah so catched on YouTube channel
I think this these stories are amazing the really fun this particular case is
simplistic talking about sort of how does the kitchen of the future a
data-driven kitchen of the future using machine learning looks like so with all
of that thank you all for being here hope that the technical sessions this
afternoon will really a picture interest and that you go home knowing more about
AWS and that you did when he walked in the door this morning so thank you all
and go build

5 Comments

Add a Comment

Your email address will not be published. Required fields are marked *